Heuristic Evaluation Function. 11. Evaluation. Issues. Admissible Evaluations. Jonathan Schaeffer

Transcription

1 Heuristic Evaluation Function. Evaluation Jonathan Schaeffer Most of the magic in a single-agent searcher is in the evaluation function To obtain an optimal answer, we need an admissible lower bound Search tree size is strongly tied to the quality of the evaluation function Unlike Alpha-beta where the evaluation function influenced the quality of the answer, but not really the size of the search //0 Issues How do we obtain an admissible evaluation function? How do we improve the quality of the evaluation function? What happens with a non-admissible heuristic? Admissible Evaluations Consider a relaxed version of the application Eliminate a rule to simplify the calculation of the heuristic distance An exact solution to a relaxed problem is usually a good heuristic for the original problem //0 //0

2 Example Consider Manhattan Distance for pathfinding Original problem Move man to goal subject to obstacles Relaxed problem Move man to goal assuming no obstacles Methodology Define the problem formally Remove one of the restrictions Evaluate whether the resulting problem is easy to evaluate and whether the results are worthwhile Could be automated, although this is an ongoing research topic //0 //0 Multiple Heuristics There may be multiple good heuristics, each of which performs better in different circumstances Given N admissible heuristics, could compose a new, more powerful heuristic: H = MAX( h, h, h, h N ) Heuristic Evaluation Right now, best way is to do this by hand; automated techniques are still in their infancy Apply application-dependent knowledge to decide on a heuristic(s) //0 //0

3 Pattern Database [] Endgame databases are a perfect lower bound for a (small) subset of positions Pattern databases computer lower bounds for subsets (patterns) of a state Using extra data, can improve the lower bound estimate Pattern Databases Define a subset of the state Enumerate all possibilities for that subset and pre-compute optimal distances to solving that relaxed problem The larger the subset the more effective the heuristic //0 //0 0 Using a Pattern Database B B Pre-compute the minimum number of moves to achieve a subset of the goal state. Using a Pattern Database Many real-world problems have symmetries that can be exploited -puzzle symmetry reflect horizontally and vertically reflect along all four axis use the maximum of all lower bounds //0 //0

4 -puzzle: H0 Simplest heuristic evaluation function Value = 0 if a goal node Value = if not a goal node -puzzle: H Count the number of misplaced tiles Assumes cost of placing a misplaced tile is 0 H = 0 H = //0 //0 -puzzle: H Manhattan Distance Count number of horizontal and vertical moves to place each tile -puzzle: H Add in linear conflicts Two tiles in a row/column that have to swap positions ( and ) 0 H = Search = 0,0 nodes 0 H = 0 //0 //0

5 -puzzle: H Pattern databases Use the pattern database shown earlier 0 H = Search (F) =,0 nodes Search(C) =, nodes Experiments () Standard set of 00 test positions Korf, A* quickly runs out of memory ( MB) can only solve the easy problems Must use iterative deepening to reduce storage needs //0 //0 Experiments () Experiments () IDA* linear storage in search depth (0 MB) Transposition table entries, 0 bytes per entry ( MB) Endgame database all positions <= moves of the goal (0 MB) Pattern database all subsets of tiles (00 MB) //0 IDA* + TT + DB + TT + DB + PDB + TT+DB+PDB,0,0, ,,,000.,,,0.,,,.,, 0.0,, fold improvement! //0 0

6 Lessons Learned Eliminate unnecessary work Any improvements to a state s evaluation will pay enormous benefits in the search space/time tradeoffs e.g. linear conflicts: a subset of pattern databases but without the storage and runtime costs Space/Time Korf has a hypothesis that there is a linear relationship between execution time and storage used Extreme cases: 0 storage and complete storage In between, roughly doubling your storage can be used to reduce the tree by roughly a factor of two //0 //0 Non-admissible Heuristics Admissible heuristics guarantee optimality Non-admissible heuristics are OK as long as you do not mind the possibility of non-optimal solutions WIDA* Multiple the h value by a small constant >.0 This has the effect of concentrating search on paths where the h value is small For many applications, this results in (near)-optimal solutions, but with a much improved execution time //0 //0

7 References [] J. Culberson and J. Schaeffer. Pattern Databases, Computational Intelligence, vol., no., pp. -,. //0